# Processing with AI
## Exploration : IA & Ethics

Name:
> Aymeric Voirin
>
Subject:
> Detect student in classe using face recognition

>[TOC]

## Design brief

### Biais

If we don't source our dataset with enough rigor, the following biais might appear:

>1. It might not work for students that are seldom represented in the data set. For example, let's say that EM Lyon wants to implement this feature and that to do so, the designer decide to use the faces of former students. They might design an AI not capable of recognising foreign student, that were only a few for the past decades but that have become an increasingly important part of EM Lyon student Body.
>2. If the Data set is choosen from a time in the paste were the student body was comprised mostly of male student, it might have trouble recognizing female students faces.

We will ensure that our model is not biaised by:

>1. Sourcing our data from various databases, for example face++ that has one of the largest database of asian faces. There's a lot of databases that exist, so I would make sure to use quite a few of them and/or select faces randomly.
>2. Making sure our data take into account skin color and tones, that male and females are equally represented in the dataset.
>3. ...

### Overfitting

We will make sure our model does not overfit by
> Checking the accuracy of our model on very different populations will be necessary. FOr example, we cannot train the model using the portraits of students of a particular year. Indeed, it won't contain enough data and won't take into account the variabilities that might occur from year to year in the student body.

### Misuse

>We have to remind ourselves that the application could be used to discriminate over classes presence instead of grades and therefore merits.
>It also poses the problem of misuse by the school and all the trust issues it could create with the student body, has it represents an intimacy intrusion against which their will be resistance and incomprehension.

### Data leakage
*Choose the most relevant proposition:*

>In a catastrophic scenario, where all of our training dataset were stolen or recovered from our model, the risk would be that people could be associated with the school they attended to. For example, in France, where diplomas are quite important, someone might not want its co-workers, for some reason, to know that he went to university and didn't go to a high standing school of business.

### Hacking

> If someone found a way to "cheat" our model and make it make any prediction that it want instead of the real one, the risk would be that the system is rendered absolutely useless. Students could trick the model and make it believe that they have been to class. FOr example, a student could be recognised twice by adding or removing an accessory. They could also find a way to be irrecognazible by the system, making it useless if most of the student decide to use the trick.